LoRA adapter (rank 64) for Qwen3.5-9B, gentle-trained on 103 QA pairs about Feb-Apr 2026 AI/ML research papers. Part of the ATLAS research architecture (bridge + specialist + text-level assembly for regulated domains).
Blind pairwise judge (Claude Opus 4.7): RAG wins 39, ATLAS wins 1, 1 tie.
This adapter plus the ATLAS architecture does NOT beat RAG-to-Opus for knowledge injection on our benchmark. It does beat Opus-without-retrieval by ~6 gold-key points. Consistent with Ovadia et al. (EMNLP 2024): RAG dominates fine-tuning for knowledge injection.
The specialist has narrow value for deployments where RAG is not viable (HIPAA / BAA scope, air-gapped, extreme query volume). It is not a replacement for RAG in general.
1from peft import PeftModel
2from transformers import AutoModelForCausalLM, AutoTokenizer
3
4base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", torch_dtype="bfloat16", device_map="auto")
5tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-9B")
6model = PeftModel.from_pretrained(base, "xpiohealth/atlas-post-cutoff-9b-specialist")
Apache 2.0 (LoRA adapter). Base model license: Qwen3.5-9B Apache 2.0.
